Driving AI Adoption and Navigating the Transition

Kevin Ortiz:

All right, good morning everyone. I’m thrilled to welcome you to our first Emerging Tech Economic Research Network event of 2026. It is so great to see a strong turnout as we kick off a new year of inquiry and discussion on the economic implications of emerging technologies such as artificial intelligence. My name is Kevin Ortiz and I serve as Co-Head of EERN here at the Federal Reserve Bank of San Francisco.

As we enter our third year of EERN, we’ve built a remarkable foundation through conversations with Nobel Prize winning economists, technology developers, leading business executives, and community groups. From academic seminars to CEO round tables, you can find all of our past events and the insights we gleaned from them on our website.

We’ll begin today’s event by tackling the rapid and uneven nature of AI adoption. To lead this conversation, I’m thrilled to welcome Fabien Curto Millet, the Chief Economist at Google, who brings a unique front row perspective on global AI implementation and economic trends. As well as Sylvain Leduc, our very own Director of Economic Research here at the San Francisco Fed. Our speakers will discuss the challenges that businesses face in transitioning to an AI environment and explore how the demand for new products will shape the jobs of the future.

As a reminder, this event is being recorded and can be accessed on our EERN website following today’s discussion. Finally, please note that the views you will hear today are those of our speakers, and do not necessarily represent the views of the Federal Reserve Bank of San Francisco or the Federal Reserve System. I’m excited for the conversation ahead, and so without further ado, Sylvain, the floor is yours.

Sylvain Leduc:

All right. Well, welcome everyone. Fabien, welcome to the San Francisco Fed. I know you’re right across the street, but it’s the first time we have you here. We’re all thrilled to have you in person, so thanks for making the trip.

Okay, so before we start, I was kind of curious about your career path. I just saw that you started your career actually having interest in central banking, monetary policy, inflation. How did you make your way to the IT sector and Google in particular?

Fabien Curto Millet:

You have done your research, Sylvain.

Sylvain Leduc:

Of course.

Fabien Curto Millet:

You take me back to the glory days of my MPhil in this field where indeed I was modeling wage equations and quantifying inflation expectations. So, all this was under the direction of John Muellbauer. And so, I was through and through a macro guy. But then what happens, an internship intervened. So, fate intervened. So this was in an economic consultancy in London that was specialized in microeconomics and especially antitrust. And I fell in love with that bit of microeconomics because it allows you to see how markets work across the economy. And so essentially, this was under Mark Williams back in London, and I just started practicing there after I completed my DPhil.

So did that for a few years, and then one bright day, Hal Varian posted a role on his team at Google for an economist. He later told me that he was looking for a “Hal” in Europe because he was traveling a little bit too much. And so that was my dream job, because it brought macroeconomics, which was sort of my university background, microeconomics, which was my background as a practicing professional economist, and computer science, which was always a sort of passion of mine. So, I’ve not looked back. This was 15 years ago and 2014 when I moved to the area and succeeded Hal two years ago.

Sylvain Leduc:

So I guess this breadth of knowledge serves you well now, right? In this role.

Fabien Curto Millet:

It does, 100%.

Sylvain Leduc:

What’s the role of this group that you have? How is this embedded in Google and how is it part of the decision making process? Is this helpful?

Fabien Curto Millet:

Yeah, no, absolutely. So, I lead a team of elite economists. They are among the creme de la creme of our field. I’m extremely proud of them. We report into the extraordinary Ruth Porat, who is President of Alphabet and Chief Investment Officer. So, we’re right at the pinnacle of a company. We look horizontally across all Alphabet activities. And so wherever there is an opportunity to sprinkle some economic goodness, we can dive in and do it.

And broadly speaking, so the work we do falls into four buckets: Number one, microeconomics, monitoring, and decoding. So, we want to understand the environment that Alphabet evolves in so that we understand its impact on our business. Number two, business economics. So this is econometrics, causal inference, just using those tools to make better business decisions. Number three, antitrust regulation policy, because with success comes scrutiny. So the lawyers do the law on these aspects and we do the economics, understanding how our conduct is efficient of the market and things of that sort.

And finally, and I guess that’s a good segue to our conversation today, thought leadership, which these days, as I say, is AI for breakfast, lunch, and dinner. So it’s every intersection of AI you can think of, AI and jobs, AI and inequality, AI and development, AI and productivity. Because as a major developer and deployer of a technology, when we have policymakers across from us, they have questions. And we better have data driven answers.

Sylvain Leduc:

All right. So, this is going to be perfect for the discussion today. Of course, there’s a lot of enthusiasm among the businesses in terms of AI. And we don’t quite see the AI adoption rate necessarily. We know there’s enthusiasm. We are doing surveys about this. Another way we can see this is the demand that companies have for complementary skills, maybe in job posting, AI related skills.

And so if you look at this, what’s interesting is you’ve seen over the years just AI generally, not only GenAI, kind of rising steadily, but now since the launch of ChatGPT around 2022, it’s really taking off a little bit more. But overall, maybe an adoption rate economy wide of about 10%, but there’s a lot of variation across sectors. The IT sector being much bigger, penetration rate of around 30% according to this measure, much less so for other sectors like retail.

So maybe a first question, just to get started, would be where do you see the biggest economic value for AI adoption? And maybe after that, some of the impediments, some of the constraints that businesses face, maybe outside of the IT sector, we’ll just abstract from the IT sector. Other sectors, what’s the biggest impediment to AI adoption?

Fabien Curto Millet:

Well, nice multi-part question. Okay. Let me try to tackle all the bits. So as you have on the chart, so there’s a lot of heterogeneity in terms of AI adoption across the economy. So I’ll mention the Census Bureau figures, which are my go to, but they’re roughly similar to what you have on the slide. So, the latest January data point was 17.7% of firms in the U.S. having adopted AI in any business function, is the sort of survey question they ask. But if you look at the information sector, that’s 35 percent-ish. If you look at construction, that’s 10% or so. So you have that kind of indeed, as you call out, that heterogeneity.

And Sundar Pichai, my CEO, expressed this a few months ago by saying that you’re going to get AJI, artificial jagged intelligence, before you get AGI, artificial general intelligence. And if you think about it, I mean, it’s not too surprising, right? I mean, AI is better at some tasks than others currently. And tasks are not uniformly distributed across the economy. So it does great at coding, information sector is all in on AI, but that’s going to sort of generalize as capabilities expand.

In terms of when you project forward the area of influence of AI, a lot of people think about it as a sort of white collar sort of thing or bicycles of the mind, as Steve Jobs used to put it. I think that that’s an overly narrow interpretation of things. Because when I step back and I look at AI, what I see is a radical reduction in the price of knowledge and analysis across the economy. And if you think of knowledge and analysis, those things are not just embedded in the production functions of white collar goods and services. They’re embedded in many, many production functions.

I’ll give you perhaps the most extreme example. So a few months ago, actually during the World Bank and IMF meetings in DC, my boss, Ruth Porat, was on stage with Ajay Banga of the World Bank. And they were unveiling the launch of open network stacks, which is a collaboration between Google and the World Bank where we bring the technology to this new platform and the World Bank brings the development expertise.
And essentially the idea was, can we bring the goodness of AI to farmers with no formal education? And they announced a pilot in Uttar Pradesh, so that pre-harvest they can decide what they plant. They can have some information around that, post-harvest, get information on market prices, where should I shift my produce? And essentially to be able to interact with that level of knowledge. So you can see we’re talking here agriculture, and this is of course a development example. Google is working to develop reliable forecasts on a 15-day basis, weather forecast. So again, influencing agriculture. So “TLDR,” I do not buy that this is a white collar phenomenon. I think it’s a whole economy phenomenon.

And I’ll come to the last part of your question, which is impediments, very briefly. So I think it was quite insightful when I looked at the management expectations survey in the UK, which is a massive survey of tens of thousands of firms across the country. They asked precisely your question. What’s kind of your blocker in terms of adopting AI? And the number one reason above even costs or lack of skills was we have not identified the business use case. And that’s not entirely surprising for a technology this powerful. You can of course deploy it very quickly as a point solution, but the big unlock comes when you deploy it as a system solution, when you reinvent your workflow. And that really takes time because it takes reinvention.

Sylvain Leduc:

Yeah, so you mentioned information, knowledge, also data. So it might not be the first thing on the line in terms of impediment, but when we talk to companies here in the region, often data security comes up. So you’re afraid of putting data out there in the public space that could be helpful to your competitors. So do you think that’s a concern and how can businesses protect their data better? Maybe even households might have this concern.

Fabien Curto Millet:

No, no, it’s a concern. So data security is, of course, especially paramount for businesses, even more paramount if you think of regulated sectors like finance. And I have a memory of McKinsey’s state of AI survey, the latest one. They were sort of asking, “What are the top things you’re trying to mitigate?” And two of the top five were indeed cybersecurity and privacy. So it’s right up there in firm’s mental models. And I suppose if you take a step back, this AI era has been opened by consumer applications, right? All these sort of chatbots that proliferated. There were, especially early on, some unsavory stories of data leakage into training sets or things of that sort. But there’s been a lot of water under the bridge. And those were consumer applications. What you now have are formalized enterprise APIs available for businesses to interact with this technology, with appropriate guardrails.

So if you come to Google Cloud, or if you really must, one of our reputable competitors, you will be looked after. And I would say that it’s a much better idea to engage with those dedicated tools as a business if you’re concerned with security, because otherwise, what you will have, is your employees are going to bring their homemade AI into your company without the sort of guardrails that you could have, and then you don’t have the same level of control. So develop a plan as a company.

Sylvain Leduc:

Yeah. And just like switching to workers a little bit, another thing for companies is ease of comfort of workers with new technology and that takes time. And I wonder in AI, because people have it on their apps and on their phone, they can just use it much more easily. You think that’s going to… like the adoption rate then will be faster as a result? Or what do companies need to do to kind of facilitate this transition?

Fabien Curto Millet:

So it’s quite funny with AI because you’ve got, again, heterogeneity here in terms of keenness for adoption. I mean, some workers are uber keen. And so you have a bring-your-own AI at work phenomenon I was referring to. McKinsey did something quite fun at the beginning of last year where they surveyed employers and surveyed employees about, “What do you think is the sort of level of use of AI in your company?” And the employers were underestimating by 3x the level of actual use that was in the company. So that’s all of your AI champions who quickly recognize that AI empowers them, removes drudgery, makes them more efficient. So you’d certainly have that. At the same time, you also find that there are many people who are hesitant. And so we ran a research study in the UK called AI Works, which engaged with workers from many sectors, education, unions, I mean, you name it.

And we found that even very simple training interventions, because it’s still early innings in AI, but it was simple interventions, doubled daily usage, and that we found that one barrier was some workers needed to feel a permission to prompt. So as a company, again, deliver those trainings, lean on them, and sort of spread the message that it’s okay to do it. So then you will bring everybody on board.

Sylvain Leduc:

So basically you have to be deliberate in your approach, no?

Fabien Curto Millet:

Precisely.

Sylvain Leduc:

Right, okay, so let’s switch to the labor market a little bit, because there’s a lot of concern. Of course, you know this. We see this in the press all the time from the whole sweats of the population about displacement of jobs from AI and maybe more a little bit on entry level jobs. We have many students often joining us online for this particular event. So what do you see about this? Is this a concern that Google has or that you have about displacing entry level jobs? And what’s happening on the ground?

Fabien Curto Millet:

No, it’s a question we pay a lot of attention to. And we’ve, in fact, done some research and I added some slides to our deck about this. So early career workers, are they hurt by AI? So lots of coverage on this topic. The best paper that makes the case that yes, there is some hurt experience at this level is Brynjolfsson and colleagues at Stanford who wrote a paper called “Canaries in the Coal Mine.” And so they think that they have spotted the sort of smoking gun evidence, if you will. The headline finding of the paper is a 16% relative decline in employment for early career workers, so at ages 22 to 25, in AI exposed occupations. So it’s a little bit of a mouthful, but that’s the encapsulation of it. And this relative decline sets in very, very quickly after the entry into the…

Sylvain Leduc:

Striking chart, really.

Fabien Curto Millet:

Right, exactly. So consumer sort of chatbots come online and then suddenly you’ve got a sort of fanning out of employment. They relied on payroll data from ADP, so the large private company, and essentially, they found this really drastic in a moment. But the timing of it seemed a little bit too neat to us. So we started doing a little bit of digging on my team. This is work with a brilliant Zanna Iscenko which we recently published via the Economic Innovation Group. And in essence, instead of looking at payroll data, we looked at job postings from Lightcast. And this is what you have on this slide, essentially. And when you look at job postings, you find that they peaked around April 2022. So six months ahead of the release of ChatGPT and big fanfare, et cetera. And so you see this fanning out behavior, we’ve sort of categorized the occupations by AI exposure quintile.

And what was happening six months ahead of ChatGPT, of course, we had inflation that needed to be brought under control. You have the highest hike in the federal funds rate in 40 years, which you have in red here. So I mean, this is the first point. The timing just doesn’t sort of stick from our perspective. It feels like a more macro driven story.

Sylvain Leduc:

So basically predating the launch of Chat… Is that the idea, like predating November 2022 and the launch of ChatGPT?

Fabien Curto Millet:

Exactly. Exactly.

Sylvain Leduc:

Okay. So more of a macro story. So the sector being sensitive to rates potentially?

Fabien Curto Millet:

That was what we hypothesized, that… Because essentially, these kind of AI exposed occupations are not randomly or uniformly distributed across the economy. If you look at the information, finance, professional, and technical services sectors, they account for 38% of employment in the highest quintile of AI exposure, but only 2 or 3% in the lowest quintile. So we figured, “Well, maybe it’s kind of interest rate sensitivity. Since then, we have a sort of interesting back and forth with Erik and co-authors who keep updating their paper and they’re a little bit skeptical about the interest rate sensitivity, but I think there is definitely some shock sensitivity that must be associated in some fashion with these AI exposure measures.

And if you move a slide, Sylvain, please. So here, let’s take it further back in time. Amusing exercise, this is the COVID-19 shock. You can see in 2020, that we have this fanning out behavior in job postings that by AI exposure quintile, except, of course, we are now well before the AI era. I mean, of course there was AI around, but it’s not like as intense as we are now. So question mark whether there is a little bit of shock correlation here with this AI exposure measure. So to be done, there are many research open goals here if you guys want to delve into this.

So, what is going on for me right now, it’s really a macro story. It’s not an easy job market by any means if you’re entering it as a sort of young person, so that is guaranteed. It’s more of a usual macro story. The last slide I think we have, Sylvain, shows you. These are job postings in high exposure AI occupations. You have one line for senior job postings and junior job postings, and you can see that they are moving in the same way. It doesn’t feel like there is discrimination targeting specifically young people. It’s just a low hire, low fire, labor market where the people who don’t have a job or are looking for one are disadvantaged.

That’s kind of the story to date. Reassurance in the present about AI does not preclude vigilance in the future. So, it should be monitored extremely carefully because again, you want to preserve these pathways to employment and shared prosperity, but that’s kind of my read of the situation.

Sylvain Leduc:

So AI doesn’t discriminate between senior and junior workers clearly?

Fabien Curto Millet:

Not on this particular evidence today.

Sylvain Leduc:

Okay. Let’s broaden the discussion a little bit on the labor market. There’s lots of concern, not only with entry level jobs, all jobs. Some pretty prominent people in the IT industries have been fairly vocal about their concerns and how much displacement we could see, pretty rapidly as a matter of fact. I sense you don’t quite share those concerns. Can you talk to us a little bit about that?

Fabien Curto Millet:

Good guess. I do not share that perspective. I mean, the correct framework to think about this in my mind is jobs lost, jobs gained, jobs changed. There’s going to be sort of activity in all of these three categories and a lot of activity in the last.

But, when I look at the lessons from history, technology is routinely a net creator of jobs. I often referenced some great work that James Manyika, my colleague who leads our technology and society division, did back when he was at the McKinsey Global Institute. He looked at the impact of personal computing rolling out across the US and essentially found that yes, it destroyed three and a half million jobs in occupations like typewriter manufacturing or clerical work, but it created 19 million positions in computer manufacturing or analysis jobs, for example. You have that dynamic with technology being introduced.

If you sort of let enough time play out, I often also quote David Autor, who’s now a technology and society fellow at Google part-time, phenomenal to work with by the way, but in his “New Frontiers” paper, he has this brilliant stat that basically 60% of employment in the US economy today is in roles that did not exist 80 years ago. That’s what you have when you let time pass.

From equilibrium to equilibrium, we are going to be just fine in my perspective. The bit to focus on, of course, with intelligent policymaking is the transition period because AI is coming at us fast. You need to deal with it responsibly. Think about your skilling, your training policies, and how you help individuals adapt to it, because the labor market loves embracing technology from cohort to cohorts, but individual adaptation is trickier. So, how do we make that happen?

Sylvain Leduc:

So, you’re concerned, I mean, I would be concerned, or I am concerned about just timing, transition for sure, but timing of job losses versus the timing of new jobs being created. Over time, yes, the new jobs are created, but you’ve got this kind of imbalance maybe along the transition. I don’t know what’s the perspective about this at Google or concerns, or we think that the new jobs, new tasks would be created faster this time around.

Fabien Curto Millet:

That goes to the dynamism of your economy. It often reminds me of that scene in Indiana Jones where he’s trying to pick up the idol and has a bag of sand and needs to do it just right. It’s kind of that moment. Where are all the new jobs going to come from, is a question I often get.

I say that that’s an interesting, but perhaps not the best question. In that simple example, bakers, there is no consumer demand for bakers. That would be cannibalistic, I don’t recommend it. There is consumer demand for bread and the demand for bakers is a derived demand. Labor demand often has that characteristic. Instead of asking where are the new jobs going to come from, maybe ask where are the new products going to come from? That gives you a completely different light on the situation.

Yes, we need labor market resilience and flexibility, and I often talk about that, but much less discussed is product market resilience and flexibility with capital market resilience and flexibility underpinning it. That’s how your economy can take turns more easily. Finally, an innovation mindset to create these new tasks and jobs that we’re discussing.

Sylvain Leduc:

All right. That’s super interesting. Let’s switch a little bit here and think about industry concentration. I mean, there’s just a lot of, AI is hungry for data. Larger companies have a lot of data. You could think a priori that would give large firms a leg up in this race. At the same time, I think when we talk to small businesses here, they’re pretty bullish on AI. They think this is going to be, for them, really helpful. Maybe new companies entering with the new technology could leapfrog older, maybe more slow to adoption for all the impediment that we discussed earlier. Is that a concern, that we’ve seen concentration rising across many industries over the past 30 years? I mean, is this something, a technology that’s just because of the interaction with data, might accentuate this, or you think that we’re more likely to see the reverse?

Fabien Curto Millet:

Time will tell as for any forward-looking statements, but my personal bet is, he who lives by the crystal ball ends up eating broken glass, Chinese proverb I often rely on. My personal bet is actually on the side of small firms, and I’ll tell you why. I go back to Ronald Coase and the theory of a firm. Why do we even have firms? Why don’t we all wake up in the morning and transact in spot markets to secure the labor and inputs that we need for production? Obviously there are efficiencies from sort of long-term contract to being able to do decision making within a firm and deploy capital, et cetera. At some point, the firm ends and markets begin, but that boundary is not set in stone. What AI does, I think, is it empowers generalists to be able to do more in more domains.

I speculate, and this is definitely speculation on my part, that we are going to favor the emergence of what I call lean corporations, a little bit along the lines of what Hal Varian and I were tracking some years ago with prior waves of digitization, where we suddenly saw the emergence of what he called micro multinationals. SMEs that are born global from day one, where you could have, one example we had was engineering in Ukraine, management in France, finance in Silicon Valley, and this was a firm that had less than 10 people. It was very, very interesting to see these kind of new kinds of animals emerge. I think there’s going to be new types of organization, AI natives that are going to emerge, and I’m going to be particularly excited to track that as well as new kinds of work that emerge across the economy.

Sylvain Leduc:

So, you think they’ll be able to reach, to get a footprint, not only like in local markets, but national, international markets to reach customers basically?

Fabien Curto Millet:

Indeed. Indeed.

Sylvain Leduc:

I see. Just in terms of maybe like if we go back to labor markets, because it’s kind of touching on this a bit, this idea that the technology helps newcomers maybe least performant, do you see this also amongst workers that once they’re coupled with AI, they become more productive and they gain even more than the more productive workers with the technology? Is that helping the more disadvantage in a sense?

Fabien Curto Millet:

I think the jury is out on that particular question and depending on which part of the literature you read, you can back up one hypothesis or the other. On the inequality reducing hypothesis or the boosting novices, you have another Eric Brynjolfsson paper on call centers where he and co-authors found that there was overall a 14% increase in issues resolved per hour thanks to AI, but it was like over 30% for novice workers. There seemed to be some sort of knowledge transmission and sort of particular boost of the novices. However, you look at other papers like Otis et al, which was about Kenyan entrepreneurs getting business recommendations from AI, and they found the opposite, that the more experienced entrepreneurs were able to make better use of the suggestions that were surfaced.

And I suppose one of the hypotheses right now in the field is maybe there is a sort of sorting task, like you need to pick among the suggestions of AI. Then experience is then a strong compliment. Whereas on narrower jobs, then actually novices benefit more. So we will see how things land. It’s likely to be fairly differentiated.

Sylvain Leduc:

It’s too new, we need more research on all those questions. So I’m cognizant of time here, but before we adjourn, I’d like to know a bit, how is the technology changing search? The way people search online, I mean, my behavior has really changed in a sense. I’m wondering, is this something like Google observes in your data? And if so, how is Google adapting to the incoming changes?

Fabien Curto Millet:

So it absolutely is a changing search. So it was changing search… I mean, it’s been doing that for years and years, but more at the backend. I mean, if you think of models we released like BERT, B-E-R-T, in terms of improvements in search ranking, et cetera. So, we were doing AI before it was cool. I mean, Sundar defined us as an AI-first company back in 2016. So we’ve been at it for a while. What’s new is you have now AI, which is more at the front end in contact directly with consumers where for example, you have these AI overviews at the top of search results.

And we have found that on queries where those AI overviews trigger, we have basically a 10% uplift in the number of queries. I mean, people really love them. Young people especially, but everybody loves them. I mean, they work really, really well. So it’s expanding the pie. And then we’ve got AI modes, which when people start using it, they use it more and more.

So Sundar in our earnings call yesterday announced that we have doubled the daily queries per user on AI mode since launch. And those queries are different. So they are three times the length of ordinary search queries. So people ask more intricate in human language questions. You’re able to ask new questions. And what I’m particularly excited about is you’re able to learn. So not just get some factual answer, but get the system to explain you well, teach me exactly why it’s the case that it is so.

And we’re seeing a lot of that behavior in our logs, which also dovetails nicely with an Ipsos survey that we released recently, which is “Our Life With AI.” 21,000 interviews across 21 countries, and we’ve seen a shift from primary use of AI as entertainment, et cetera, to uses for learning. So that was the message of the survey. And one I’m really excited to track as we go forward.

Sylvain Leduc:

Okay, last question. You’re right there at the forefront of technology. We’re not, but we’re really curious. So what do we have to be on the lookout for? What’s really exciting people at Google right now?

Fabien Curto Millet:

Well, I hope nobody has made dinner plans. Okay. Let me give you a few. AI and science. I am very excited about AI and science. AI is turbocharging scientific discovery. Take AlphaFold. We release to the world the structure of 200 million proteins that the scientific community is now using to work on malaria treatments, cancer treatments. And bear in mind that before this system, cracking the structure of a single protein could take years in the lab and cost hundreds of thousands of dollars. Just imagine the productivity uplift. So AI and science is turbocharged, which will show up at some point in our TFP (total factor productivity) statistics, I’m hoping. That’s one. AI and public sector. The public sector is what, 15% of employment across OECD economies, and it’s a key complement to the whole private sector economy. And whether it’s fixing potholes in Memphis or digitizing all planning documents in the UK, so smoothing the permitting process, AI is accelerating all of that.

AI and education, personalized education is the next frontier. We were talking about learning, so I’m very excited about that. And I’ll finish, since we’re in negative time, on AI and health, where AI is empowering workers. So nurses, nurse practitioners are doing more work with higher expertise. It’s removing drudgery. And I think President Daly often mentioned this, that people don’t get into the healthcare sector for paperwork. They get in the healthcare sector for care and they do preciously little of that sometimes. So solving that burnout issue. And early treatment, early detection of say cancers. So AI is helping on all of those fronts and revolutionizing healthcare.

Sylvain Leduc:

Thank you so much for making the short walk and sharing your perspectives with us today. This has been a fascinating discussion. For those of you who are interested, our next event will be with Nick Bloom who’s going to talk a little bit more about AI adoption rates. So I really hope you can join us then. Fabien, thank you so much for being with us today.

Fabien Curto Millet:

Thank you, Sylvain. Thanks.